Abstract
dc:description.abstractStatistical methodology plays a vital role in understanding and interpreting complex datasets, particularly in the context of modern biological research. The increasing availability of high-dimensional and multi-way data presents both opportunities and challenges for researchers. Effective analysis of such data requires innovative approaches capable of integrating diverse data sources and addressing their inherent complexities. This dissertation focuses on three statistical methodologies that address these challenges: the integrative analysis of multiple multi-way datasets, mediation analysis involving high-dimensional omics data and tensor mediation analysis using latent models. The first contribution introduces a penalized framework for integrative analysis of multiple linked tensors. Motivated by studies where multi-way datasets share common subject modes but differ in dimensional structure, the proposed method enables joint factorization while automatically identifying shared and modality-specific latent components. The framework accommodates heterogeneous tensor dimensions and structured missingness, providing an interpretable low-rank representation in multi-modal settings. The second contribution addresses mediation analysis with high-dimensional omics mediators under case-control study designs. Classical mediation approaches are unstable when the number of mediators is large and effects are weak. To overcome these limitations, a variance-based total mediation effect measure with causal interpretation is developed, along with a cross-fitted estimation procedure that remains robust under high dimensionality. The proposed method provides stable and interpretable quantification of the overall mediating role of molecular profiles in exposure–disease relationships. The third contribution integrates tensor modeling with mediation analysis to address settings in which mediators themselves possess multi-way structure. Motivated by proteomics data organized as subject-by-time-by-protein arrays, a latent tensor framework is proposed to achieve dimension reduction while preserving the total mediation effect under transformation. This approach enables computationally feasible and statistically valid inference on the mediating role of tensor-structured biomarkers. Together, these contributions advance statistical methodology for high-dimensional and multi-way data, providing scalable tools for uncovering latent structure and elucidating biological mechanisms in modern biomedical research.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kang, Zhiyu
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Kang, Zhiyu. (2026). High-dimensional Tensor Decomposition and Mediation Analysis. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/280368.